Plain-English guide
AI website assistants
for small businesses.
Written by someone who builds them, which means I know exactly where they fall over. Here is what these things actually do, the failure nobody selling one will mention, and how to work out whether one would earn its keep in your business.
An AI website assistant answers visitor questions from your own business information and captures the enquiry when someone is ready to buy. It is worth it if you lose enquiries outside working hours or answer the same questions repeatedly. The main risk is that most assistants are set up once and go quietly out of date, confidently quoting prices and services that have changed.
Published
What one actually does
Forget the chatbots you remember. Those followed a decision tree — press 1 for sales — and everybody hated them, rightly. A modern assistant answers from your business's own information: your services, your prices, your process, the questions you get asked forty times a week.
The sequence is short. Somebody lands on your site at nine at night. They have one question standing between them and enquiring — do you cover my area, do you do that type of work, roughly what does it cost. The assistant answers it. Then, because they are now interested rather than wondering, it takes their details and sends you the enquiry.
Note what it is not doing. It is not closing a sale, quoting a job, or replacing you. It is removing the reason somebody clicks off your site and onto the next one.
Whether it is worth it for you
Three questions, answered honestly. If all three are no, do not buy one — from me or anybody.
Do people visit your site outside working hours?
Look at your analytics rather than guessing. Evenings and weekends are usually a bigger share than owners expect, and those are the visits where nobody is there to answer.
Do you answer the same questions constantly?
Areas covered, lead times, whether you do that type of job, how you quote. If you could recite them, an assistant handles them and gives you the time back.
Do enquiries dry up between interest and contact?
Plenty of traffic, few enquiries, usually means people had a question your site did not answer and they went elsewhere rather than asking.
If your work comes entirely by referral and nobody browses before ringing, an assistant is solving a problem you do not have. I would rather tell you that than sell you one.
The bit nobody mentions
They go quietly out of date
This is the real failure mode, and you will not hear it from anyone selling you an assistant, because it happens six months after the invoice is paid.
Here is the pattern. Somebody sets an assistant up, loads it with your services and prices as they were that week, and puts it live. It works beautifully and everyone is pleased. Then the business moves on, as businesses do. Prices change. You drop a service. Lead times shift. A supplier changes. Nobody goes back to the assistant, because it is not broken and nothing is prompting anyone to.
A year later it is confidently telling your customers things that stopped being true in the spring.Confidently wrong is worse than obviously broken, because nobody reports it, you never find out, and you may be held to what it said.
So when you are comparing assistants, the question that matters is not how clever it sounds in a demo. It is: how does this thing stay right? What happens when a price changes? How would I know if it answered something wrongly last Tuesday? If the answer is "you'd have to check it yourself," you have bought a maintenance job you will not do.
You are liable for what it says
Worth being blunt about. If an assistant on your website quotes a customer a price, that is your business quoting a price. If it tells them you can be there Thursday, that is your commitment. "The AI said it" is not a position you want to be arguing from.
The risk is not theoretical. An assistant running on a general model, with no grounding in what you actually charge, will produce a plausible-sounding figure rather than admitting it does not know. It is not lying; it is doing what it was built to do, which is generate a likely-looking answer.
The fix is a build decision. An assistant that only answers from your real information, and hands over rather than guessing, does not do this. "I don't want to give you a wrong figure, so I'll get Mick to call you back" is a better answer than an invention, and customers prefer it too.
What separates a good one from an annoying one
Everyone has been trapped by a bad one. The differences are not subtle.
It knows your business, not your category
Trained on your services, your areas, your prices, your actual FAQs. Generic assistants give generic answers and visitors can tell within one exchange.
The escape hatch is offered, not hidden
People forgive an assistant that cannot answer something. They do not forgive being stuck with it. Offering a person early is what prevents complaints.
It says what it is
Before they type. Visitors who know they are talking to an assistant ask clearer questions and get better results. Visitors who work it out halfway through feel deceived.
It hands over rather than guesses
And it tells you what it could not answer, so the gap gets filled rather than repeating for the next hundred visitors.
It handles your customers' data properly
It is collecting names, numbers and job details. That is personal data, the provider is processing it on your behalf, and it needs to be built and hosted to a standard you would be comfortable explaining.
The legal bits
Two duties, both straightforward, both routinely missed.
Say it is AI. No single UK statute requires it, though consumer protection law is unfriendly to a bot posing as a person, and it has been a legal requirement since August 2026 if any of your customers are in the EU.The full picture is here.
Update your privacy notice. The assistant collects personal data and whoever provides it processes that data on your behalf. That is a separate duty from the disclosure, and firms routinely do one and miss the other.
Go and try a real one
Reading about these is less useful than using one. Cornwall Play Safety, an RPII-accredited playground inspection firm, run an assistant called Piran on their site. It answers questions about inspection tiers and report turnaround, and passes anything site-specific to the inspector rather than guessing.
Go and try to catch it out. Ask it something it should not know. That is the honest test, and it is more informative than any demo built to impress you.
Try Piran live ↗What to do this week
- 01
Check when people visit. Your analytics will tell you what share of traffic arrives outside working hours. That number is the size of the problem.
- 02
Write down your five most-asked questions. If you can do it from memory in two minutes, an assistant can handle them.
- 03
If you already have one, ask it three awkward questions. A price, availability, and something it should not know. If it invents answers, that is a bigger problem than anything on your site.
- 04
Ask any supplier how theirs stays accurate. Not how clever it is. How it stays right in a year. The answer tells you what you are actually buying.
Straight answers
- What is an AI website assistant?
- Software that sits on your website and holds a short conversation with visitors — answering questions about your services, prices and process, then taking the enquiry when someone is ready. Unlike an old-fashioned chatbot it is not following a decision tree; it answers from your business’s own information.
- Is it worth it for a small business?
- It depends on whether you lose enquiries. If most visitors arrive outside working hours, or you get the same five questions all day, it earns its keep quickly. If your work is entirely word of mouth and nobody browses your site before ringing, it will not.
- Will it annoy my customers?
- A bad one will. The things that irritate people are being trapped with no route to a person, being answered with waffle, and working out halfway through that it is a bot. All three are build decisions, not inevitable properties of AI.
- What is the biggest problem with these assistants?
- They go quietly out of date. Most are loaded up once at set-up and never revisited, so a year later they are confidently quoting old prices and services you dropped. That is worse than being broken, because nobody reports it and you can be held to what it said.
- Am I liable for what it tells people?
- Yes. A price it invents is a price you may have to honour, and an assurance it gives is one you gave. This is the strongest argument for grounding an assistant in your real services rather than letting a general model improvise.
- Do I have to tell visitors it is AI?
- It should say so, upfront. There is no single UK statute requiring it, though consumer protection law is unfriendly to a bot posing as a person, and it has been a legal duty since August 2026 if you serve EU customers. It also works better — people who know adjust how they ask.
- What does one cost?
- Typically a set-up fee plus a monthly. The monthly is not just hosting — it is what keeps the assistant accurate as your business changes. An assistant nobody maintains is worse than no assistant, so treat a suspiciously cheap one with caution.
- How is this different from a contact form?
- A form asks people to do work and wait. An assistant answers the question that was stopping them, then captures the enquiry as a by-product. Most people who would have bounced off a form will talk to an assistant.
Read next
Do I have to tell customers they are talking to a bot?
No single UK law says so. EU law now does. Where the line sits, and why one sentence solves it either way.
Is my business ready for AI? A ten-point check
Score yourself in ten minutes. Where you already stand, what is missing, and what is worth fixing first.
Want one that stays right?
Foredesk is built around the staleness problem: every answer carries a date, anything it could not answer comes back to you, and information that has gone out of date gets retired rather than repeated. Trained on your services, in your voice, with disclosure and a route to a person as standard.
Ask about Foredesk